Job Summary

Job Description — AI/ML Engineer (Codibix Assistant)

Location: Chennai

Experience level: Senior — 8–12+ years

Summary

Owns the conversational Assistant end-to-end — a standalone Python/Node service, deliberately separate from the Java/Spring Boot core, that authors and validates knowledge-repository entries through a multi-agent architecture. This is the one genuinely novel, non-boilerplate part of the platform: not a CRUD app Claude can generate wholesale, but a governed agent system where every AI-drafted output must stay clearly distinguishable from human-authored content until a Reviewer approves it.

Key Responsibilities

Design and build the multi-agent architecture: Orchestrator/Router, Retrieval/Grounding (RAG), Authoring, Validation/Schema, and Explainer/Audit agents — each with a narrow, explicit job, not one do-everything LLM call

Enforce the core guardrail in code, not just in the prompt: the Authoring agent drafts only, never auto-publishes — every AI-suggested entry must render visibly distinct from human-authored ones

Build the tool-calling contract per agent (e.g., Authoring agent can call draft_entry() but never publish_entry()) — RBAC boundaries enforced at the tool layer, not hoped for in the system prompt

Own the retrieval layer: embedding strategy, chunking, and query design against the shared PostgreSQL/pgvector store so drafts don't duplicate or contradict existing repository content

Build and maintain the model-agnostic integration layer — the abstraction that makes "swap the underlying model, keep the same agent behavior" actually true in code, not just a slide claim

Own the REST/gRPC API contract this service exposes to the Java core (owned jointly with the Solution Architect, who owns the consuming side of that same contract)

Define and monitor agent-quality metrics — hallucination rate on drafted entries, retrieval relevance, false-approval risk on the Validation agent

Partner with Domain Architects to validate that agent-drafted entries actually reflect real domain judgment, not plausible-sounding text

Skill Requirements

5–8 years in ML/AI engineering, with direct experience building production LLM-based agent systems, not just prototype chatbots

Hands-on experience with an agent orchestration framework (e.g., LangGraph, LangChain, or equivalent) and a clear point of view on why a graph/state-machine approach beats a single unstructured agent loop for governed use cases

Strong Python or Node.js (matches whichever the Assistant service is ultimately built in)

Experience with retrieval-augmented generation: embeddings, vector search (pgvector or equivalent), chunking strategy trade-offs

Demonstrated ability to design guardrails that are structurally enforced, not prompt-only — this is the single most important skill for this role, given the platform's audit-trail requirements

Comfortable owning a service boundary/API contract with a team building in a different language (Java) — this role does not need Java fluency, but does need to design a contract a Java consumer can integrate against cleanly

Experience evaluating LLM output quality in production (not just offline benchmarks) — hallucination detection, relevance scoring, human-feedback loop design

Other Requirements

Preferred                

Prior experience in a regulated or compliance-sensitive domain where AI output requires traceability back to source

Familiarity with Spring AI or LangChain4j (useful context for the API contract, even though this role doesn't build in Java)

Experience building a correction/feedback loop that feeds structured signal back into a knowledge base, not just a thumbs-up/down

Success Metrics

Draft-to-approval rate (how often Reviewers accept vs. reject/heavily edit AI-drafted entries) — the most direct signal of whether the Assistant is actually helping

Retrieval relevance (drafts genuinely grounded in existing repository content, not duplicated or contradicted)